-
Notifications
You must be signed in to change notification settings - Fork 46
/
main.py
262 lines (203 loc) · 11.5 KB
/
main.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
import os
import argparse
from itertools import chain
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.autograd import Variable
from utils import *
from model import *
from download import CelebA
import scipy
import imageio
from progressbar import ETA, Bar, Percentage, ProgressBar
parser = argparse.ArgumentParser(description='PyTorch implementation of DiscoGAN')
parser.add_argument('--cuda', type=str, default='true', help='Set cuda usage')
parser.add_argument('--epoch_size', type=int, default=5000, help='Set epoch size')
parser.add_argument('--batch_size', type=int, default=64, help='Set batch size')
parser.add_argument('--learning_rate', type=float, default=0.0002, help='Set learning rate for optimizer')
parser.add_argument('--result_path', type=str, default='./results/', help='Set the path the result images will be saved.')
parser.add_argument('--model_path', type=str, default='./models/', help='Set the path for trained models')
parser.add_argument('--image_size', type=int, default=64, help='Image size. 64 for every experiment in the paper')
parser.add_argument('--gan_curriculum', type=int, default=10000, help='Strong GAN loss for certain period at the beginning')
parser.add_argument('--starting_rate', type=float, default=0.01, help='Set the lambda weight between GAN loss and Recon loss during curriculum period at the beginning. We used the 0.01 weight.')
parser.add_argument('--default_rate', type=float, default=0.5, help='Set the lambda weight between GAN loss and Recon loss after curriculum period. We used the 0.5 weight.')
parser.add_argument('--style_A', type=str, default=None, help='Style for CelebA dataset. Could be any attributes in celebA (Young, Male, Blond_Hair, Wearing_Hat ...)')
parser.add_argument('--style_B', type=str, default=None, help='Style for CelebA dataset. Could be any attributes in celebA (Young, Male, Blond_Hair, Wearing_Hat ...)')
parser.add_argument('--constraint', type=str, default=None, help='Constraint for celebA dataset. Only images satisfying this constraint is used. For example, if --constraint=Male, and --constraint_type=1, only male images are used for both style/domain.')
parser.add_argument('--constraint_type', type=str, default=None, help='Used along with --constraint. If --constraint_type=1, only images satisfying the constraint are used. If --constraint_type=-1, only images not satisfying the constraint are used.')
parser.add_argument('--n_test', type=int, default=200, help='Number of test data.')
parser.add_argument('--update_interval', type=int, default=3, help='')
parser.add_argument('--log_interval', type=int, default=50, help='Print loss values every log_interval iterations.')
parser.add_argument('--image_save_interval', type=int, default=1000, help='Save test results every image_save_interval iterations.')
parser.add_argument('--model_save_interval', type=int, default=10000, help='Save models every model_save_interval iterations.')
parser.add_argument('--attrs', '--list', nargs='+', help='selected attributes for the CelebA dataset',
default=['Black_Hair', 'Blond_Hair', 'Male'])
args = parser.parse_args()
import torchvision.transforms as transforms
from torchvision.datasets.vision import VisionDataset
transform=transforms.Compose([
transforms.RandomHorizontalFlip(p=0.5),
transforms.CenterCrop(178),
transforms.Resize(size=64),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5),(0.5, 0.5, 0.5))
])
dataset=CelebA(root='./',attributes=args.attrs,transform=transform,download=True)
data_A, data_B = get_celebA_files(style_A=args.style_A, style_B=args.style_B, constraint=args.constraint, constraint_type=args.constraint_type, test=False, n_test=args.n_test)
test_A, test_B = get_celebA_files(style_A=args.style_A, style_B=args.style_B, constraint=args.constraint, constraint_type=args.constraint_type, test=True, n_test=args.n_test)
def as_np(data):
return data.cpu().data.numpy()
def get_fm_loss(real_feats, fake_feats, criterion):
losses = 0
for real_feat, fake_feat in zip(real_feats, fake_feats):
l2 = (real_feat.mean(0) - fake_feat.mean(0)) * (real_feat.mean(0) - fake_feat.mean(0))
loss = criterion( l2, Variable( torch.ones( l2.size() ) ).cuda() )
losses += loss
return losses
def get_gan_loss(dis_real, dis_fake, criterion, cuda):
labels_dis_real = Variable(torch.ones( [dis_real.size()[0], 1] ))
labels_dis_fake = Variable(torch.zeros([dis_fake.size()[0], 1] ))
labels_gen = Variable(torch.ones([dis_fake.size()[0], 1]))
if cuda:
labels_dis_real = labels_dis_real.cuda()
labels_dis_fake = labels_dis_fake.cuda()
labels_gen = labels_gen.cuda()
dis_loss = criterion( dis_real, labels_dis_real ) * 0.5 + criterion( dis_fake, labels_dis_fake ) * 0.5
gen_loss = criterion( dis_fake, labels_gen )
return dis_loss, gen_loss
cuda = args.cuda
if cuda == 'true':
cuda = True
else:
cuda = False
epoch_size = args.epoch_size
batch_size = args.batch_size
result_path = args.result_path
if args.style_A:
result_path = os.path.join( result_path, args.style_A )
model_path = args.model_path
if args.style_A:
model_path = os.path.join( model_path, args.style_A )
data_style_A, data_style_B, test_style_A, test_style_B = data_A, data_B, test_A, test_B
test_A = read_images( test_style_A, None, args.image_size )
test_B = read_images( test_style_B, None, args.image_size )
test_A = Variable( torch.FloatTensor( test_A ), volatile=True )
test_B = Variable( torch.FloatTensor( test_B ), volatile=True )
if not os.path.exists(result_path):
os.makedirs(result_path)
if not os.path.exists(model_path):
os.makedirs(model_path)
generator_A = Generator()
generator_B = Generator()
discriminator_A = Discriminator()
discriminator_B = Discriminator()
if cuda:
test_A = test_A.cuda()
test_B = test_B.cuda()
generator_A = generator_A.cuda()
generator_B = generator_B.cuda()
discriminator_A = discriminator_A.cuda()
discriminator_B = discriminator_B.cuda()
data_size = min( len(data_style_A), len(data_style_B) )
n_batches = ( data_size // batch_size )
recon_criterion = nn.MSELoss()
gan_criterion = nn.BCELoss()
feat_criterion = nn.HingeEmbeddingLoss()
gen_params = chain(generator_A.parameters(), generator_B.parameters())
dis_params = chain(discriminator_A.parameters(), discriminator_B.parameters())
optim_gen = optim.Adam( gen_params, lr=args.learning_rate, betas=(0.5,0.999), weight_decay=0.00001)
optim_dis = optim.Adam( dis_params, lr=args.learning_rate, betas=(0.5,0.999), weight_decay=0.00001)
iters = 0
gen_loss_total = []
dis_loss_total = []
for epoch in range(epoch_size):
data_style_A, data_style_B = shuffle_data( data_style_A, data_style_B)
widgets = ['epoch #%d|' % epoch, Percentage(), Bar(), ETA()]
pbar = ProgressBar(maxval=n_batches, widgets=widgets)
pbar.start()
for i in range(n_batches):
pbar.update(i)
generator_A.zero_grad()
generator_B.zero_grad()
discriminator_A.zero_grad()
discriminator_B.zero_grad()
A_path = data_style_A[ i * batch_size: (i+1) * batch_size ]
B_path = data_style_B[ i * batch_size: (i+1) * batch_size ]
A = read_images( A_path, None, args.image_size )
B = read_images( B_path, None, args.image_size )
A = Variable( torch.FloatTensor( A ) )
B = Variable( torch.FloatTensor( B ) )
if cuda:
A = A.cuda()
B = B.cuda()
AB = generator_B(A)
BA = generator_A(B)
ABA = generator_A(AB)
BAB = generator_B(BA)
# Reconstruction Loss
recon_loss_A = recon_criterion( ABA, A )
recon_loss_B = recon_criterion( BAB, B )
# Real/Fake GAN Loss (A)
A_dis_real, A_feats_real = discriminator_A( A )
A_dis_fake, A_feats_fake = discriminator_A( BA )
dis_loss_A, gen_loss_A = get_gan_loss( A_dis_real, A_dis_fake, gan_criterion, cuda )
fm_loss_A = get_fm_loss(A_feats_real, A_feats_fake, feat_criterion)
# Real/Fake GAN Loss (B)
B_dis_real, B_feats_real = discriminator_B( B )
B_dis_fake, B_feats_fake = discriminator_B( AB )
dis_loss_B, gen_loss_B = get_gan_loss( B_dis_real, B_dis_fake, gan_criterion, cuda )
fm_loss_B = get_fm_loss( B_feats_real, B_feats_fake, feat_criterion )
# Total Loss
if iters < args.gan_curriculum:
rate = args.starting_rate
else:
rate = args.default_rate
gen_loss_A_total = (gen_loss_B*0.1 + fm_loss_B*0.9) * (1.-rate) + recon_loss_A * rate
gen_loss_B_total = (gen_loss_A*0.1 + fm_loss_A*0.9) * (1.-rate) + recon_loss_B * rate
gen_loss = gen_loss_A_total + gen_loss_B_total
dis_loss = dis_loss_A + dis_loss_B
if iters % args.update_interval == 0:
dis_loss.backward()
optim_dis.step()
else:
gen_loss.backward()
optim_gen.step()
if iters % args.log_interval == 0:
print ("---------------------")
print ("GEN Loss:", as_np(gen_loss_A.mean()), as_np(gen_loss_B.mean()))
print ("Feature Matching Loss:", as_np(fm_loss_A.mean()), as_np(fm_loss_B.mean()))
print ("RECON Loss:", as_np(recon_loss_A.mean()), as_np(recon_loss_B.mean()))
print ("DIS Loss:", as_np(dis_loss_A.mean()), as_np(dis_loss_B.mean()))
if iters % args.image_save_interval == 0:
AB = generator_B( test_A )
BA = generator_A( test_B )
ABA = generator_A( AB )
BAB = generator_B( BA )
n_testset = min( test_A.size()[0], test_B.size()[0] )
subdir_path = os.path.join( result_path, str(iters / args.image_save_interval) )
if os.path.exists( subdir_path ):
pass
else:
os.makedirs( subdir_path )
for im_idx in range( n_testset ):
A_val = test_A[im_idx].cpu().data.numpy().transpose(1,2,0) * 255.
B_val = test_B[im_idx].cpu().data.numpy().transpose(1,2,0) * 255.
BA_val = BA[im_idx].cpu().data.numpy().transpose(1,2,0)* 255.
ABA_val = ABA[im_idx].cpu().data.numpy().transpose(1,2,0)* 255.
AB_val = AB[im_idx].cpu().data.numpy().transpose(1,2,0)* 255.
BAB_val = BAB[im_idx].cpu().data.numpy().transpose(1,2,0)* 255.
filename_prefix = os.path.join (subdir_path, str(im_idx))
imageio.imwrite( filename_prefix + '.A.jpg', A_val.astype(np.uint8)[:,:,::-1])
imageio.imwrite( filename_prefix + '.B.jpg', B_val.astype(np.uint8)[:,:,::-1])
imageio.imwrite( filename_prefix + '.BA.jpg', BA_val.astype(np.uint8)[:,:,::-1])
imageio.imwrite( filename_prefix + '.AB.jpg', AB_val.astype(np.uint8)[:,:,::-1])
imageio.imwrite( filename_prefix + '.ABA.jpg', ABA_val.astype(np.uint8)[:,:,::-1])
imageio.imwrite( filename_prefix + '.BAB.jpg', BAB_val.astype(np.uint8)[:,:,::-1])
if iters % args.model_save_interval == 0:
torch.save( generator_A, os.path.join(model_path, 'model_gen_A-' + str( iters / args.model_save_interval )))
torch.save( generator_B, os.path.join(model_path, 'model_gen_B-' + str( iters / args.model_save_interval )))
torch.save( discriminator_A, os.path.join(model_path, 'model_dis_A-' + str( iters / args.model_save_interval )))
torch.save( discriminator_B, os.path.join(model_path, 'model_dis_B-' + str( iters / args.model_save_interval )))
iters += 1